1. Consequence driven decomposition
of large-scale power system security
analysis
Florence Fonteneau-Belmudes∗, Damien Ernst∗,
Christophe Druet†, Patrick Panciatici‡, Louis
Wehenkel∗
*University of Li`ege, Belgium - † ELIA (Belgian TSO) - ‡ RTE (French TSO)
2010 IREP Symposium, Buzios, Rio de Janeiro, Brazil, 1-6 August 2010

2. Power system security assessment
Power system security assessment = Identifying
contingencies/scenarios that could lead to unacceptable
operating conditions (dangerous contingencies) if no
preventive actions were taken.
Problem may be approximated by running a security
analysis for every element of a set of potentially
dangerous contingencies.

3. Size of this set of potentially dangerous contingencies:
• grows with the size of the power system
• grows with increasing uncertainties about the future
power injection patterns, load patterns and system
topologies.
Question: What should you do when the size of the set
becomes so large that you cannot screen every
contingency one by one?
2

4. Traditional solutions:
• Increase the computational resources (note that the
security analysis task can be easily parallelized).
• Use filtering techniques (= simplified conservative
analysis techniques) to identify a subset of
contingencies on which the full security analysis is
carried out.
Shortcomings:
• Do not work in case of very large number of
contingencies (say for example 1010).
• Reliable and efficient filtering techniques are difficult
to develop.
3

5. Our solution
A generic approach to identify dangerous contingencies
by running a full security analysis for a number of
contingencies greater that the number of dangerous ones
but much smaller than the number of potentially
dangerous contingencies.
Why generic? Work whatever the type of security
analysis (e.g., static security, transient stability, voltage
stability).
Come atop of existing security analysis tools and can be
combined with any of them.
4

6. And no magic is necessary for the approach to work...
5

7. How does it work:
1. Draw at random a subset of contingencies C.
2. Run a security analysis for every element of C. Add
the dangerous contingencies to the set DC.
3. Use the information collected from the security
analyses to draw a new set C more likely to contain
new dangerous contingencies.
4. If computational resources are exhausted output DC,
else go back to 2.
6

8. 3. Use the information collected from the security analyses to draw a new
set C more likely to contain new dangerous contingencies.
Tools needed:
• A post-processor of the security analysis results to
associate to a contingency a severity index (e.g., amount
of load lost in the post-fault configuration).
• A method to embed the contingency space in a compact
metric space where the severity indices vary in a
“progressive way”.
Solution:
1. Identify among the contingencies already analyzed those
with the highest severity index.
2. Identify among a family of sampling densities over the
compact metric space, the one which is the most likely to
generate the contingencies identified at Step 1.
3. Draw a new set C using the sampling density identified at
Step 2.
7

9. Illustrative example
Test network:
Set of potentially dangerous contingencies: N-1
contigencies, loss of a transmission element, 634 elements.
Available computational resources: 125 security analyses
(load-flows) can be carried out.
What we want: To identify the 6 contingencies that cause an
overload on a critical transmission line.
8

10. Tools needed:
• A post-processor of the security analysis results to associate to a
contingency a severity index (e.g., amount of load lost in the post-fault
configuration).
• A method to embed the contingency space in a compact metric space
where the severity indices vary in a “progressive way”.
• Severity index: overload on the critical transmission line.
• Compact metric space: the plane is divided into
surfaces corresponding to the contingencies according to
their location on the geographical map.
9

11. A typical run of the algorithm with multi-dimensional
Gaussian sampling densities:
Iteration 1 Iteration 2
Iteration 3 Iteration 4
10

12. Probability of finding at least
n dangerous contingencies.
n Our approach Monte Carlo
1 0.99 0.69
2 0.90 0.31
3 0.76 0.09
4 0.47 0.04
5 0.20 0
6 0.04 0
Comments:
• Results get even better with the decrease of the ratio:
number of dangerous contingencies
number of potentially dangerous contingencies
• Many improvements have already been brought to the
basic version of the approach presented here.
11

13. Willing to know more...
“Consequence driven decomposition of large-scale power system security
analysis”. F. Fonteneau-Belmudes, D. Ernst, C. Druet, P. Panciatici and L.
Wehenkel. In Proceedings of the 2010 IREP Symposium - Bulk Power
Systems Dynamics and Control - VIII, Buzios, Rio de Janeiro, Brazil, 1-6
August 2010. (8 pages).
“Pseudo-geographical representations of power system buses by
multidimensional scaling”. F. Belmudes, D. Ernst and L. Wehenkel. In
Proceedings of the 15th International Conference on Intelligent System
Applications to Power Systems (ISAP 2009), Curitiba, Brazil, 8-12 November
2009. (6 pages).
“A rare-event approach to build security analysis tools when N − k (k > 1)
analyses are needed (as they are in large-scale power systems)”. F. Belmudes,
D. Ernst and L. Wehenkel. In Proceedings of the 2009 IEEE Bucharest Power
Tech Conference, June 28th-July 2nd, Bucharest, Romania. (8 pages).
“Cross-entropy based rare-event simulation for the identification of dangerous
events in power systems”. F. Belmudes, D. Ernst and L. Wehenkel. In
Proceedings of the 10th International Conference on Probabilistic Methods
Applied to Power Systems (PMAPS-08), pages 1290-1295. Rincon, Puerto
Rico, 25-29 May 2008.
12